Computer Vision Engineer / Applied Mathematician

Location

Hyderabad — in office all 5 days, as the role requires substantial brainstorming within the team.

Salary range

₹3–6 LPA, depending on the candidate's suitability for the position

Position start date

1 December 2026

Project Overview

We are developing a high-precision, post-capture video analytics engine for high-speed dynamic objects using a lightweight camera-to-laptop asynchronous processing pipeline. We are seeking a Computer Vision Engineer / Applied Mathematician to build core spatial math algorithms and feature tracking pipelines.

The target system tracks custom geometric surface features on spherical bodies captured at high frame rates (200+ FPS) with fast shutter speeds.

Key Responsibilities & Scope of Work

1. Asymmetric Pattern Detection & Contour Isolation Network
   • Build a robust OpenCV / PyTorch segmentation pipeline to extract the 3 vertices of custom asymmetric hollow triangles from 240 FPS video frames.
   • Implement morphological operations (cv2.MORPH_CLOSE, cv2.MORPH_OPEN) and adaptive thresholding to filter out ISO 1600 sensor noise and minor lighting reflections.
   • Maintain deterministic vertex tracking across extreme rotational motion blurs (Δt = 4.16 ms frame step).
2. Perspective-n-Point (PnP) & Quaternion 3D Rotation Solver
   • Formulate and implement a Perspective-n-Point (PnP) rotation engine (cv2.solvePnP / cv2.solvePnPRansac) mapping 2D image coordinates to a 3D spherical model (40mm diameter).
   • Compute continuous frame-to-frame 3D rotation vectors (q = w + xi + yj + zk) to resolve full 3-axis angular spin vectors up to 100 RPS / 6,000 RPM.
   • Ensure rotational ambiguity resolution using asymmetric geometric constraints.
3. CMOS Rolling Shutter Affine Shear Compensation
   • Develop an affine shear un-warp algorithm compensating for the GoPro Hero 12 rolling shutter delay (0.26ms line readout).
   • Model the temporal spatial shift matrix to correct 2D bounding-box geometric shearing on high-speed rotating objects prior to 3D PnP execution.

Required Technical Qualifications

• Core Expertise: Advanced proficiency in Python, OpenCV, NumPy, SciPy, and C++ (optional, for acceleration).
• Mathematical Background: Deep understanding of 3D Projective Geometry, Camera Intrinsic Calibration Matrices (K), Quaternion Algebra, Homography, and PnP math solvers.
• Sensor Experience: Hands-on experience working with high-speed video processing (240+ FPS), rolling shutter distortion matrices, and camera distortion models (k1, k2, p1, p2).
• Deliverable Standards: Clean, well-documented, unit-tested Python modules ready for integration into a desktop ingestion engine.

Desirable / Preferred Qualifications

• Synthetic Data & Blender Pipeline Familiarity: Experience using Blender (or similar 3D engines) to extract ground-truth camera matrices, 3D target coordinates, or rendered synthetic frame sequences for computer vision validation.
• C++ / PyBind11 Acceleration: Ability to convert computational bottlenecks (such as high-speed frame-by-frame PnP matrix calculations or un-warp loops) into compiled C++ extensions or CUDA kernels if Python throughput needs optimization later.
• PyTorch / Deep Learning Tracking: Experience with lightweight neural network trackers (e.g., YOLOv8-Pose, MobileNet-V3) or custom landmark/keypoint detection models for feature extraction under heavy occlusion or motion blur.
• Physics & Kinematics Modeling: Basic background in classical rigid-body dynamics (calculating angular velocity vectors, translational trajectories, or Magnus effect drag modeling) to sanity-check extracted spin values.
• Desktop Application Ingestion Experience: Familiarity with batch video ingestion pipelines, multi-threading (concurrent.futures), or asynchronous IO processing for handling large video file payloads efficiently.

Deliverables & Acceptance Criteria

• Module 1: Contour Pipeline — Python module for asymmetric triangle extraction. Acceptance: ≥ 98% target vertex detection rate under ISO 1600 noise.
• Module 2: Rolling Shutter Shear Matrix — Affine compensation script for GoPro Hero 12 down-scanning lines (0.26ms). Acceptance: Reconstructs true circular aspect ratios (≥ 0.95 circularity score) at 2,200+ RPM.
• Module 3: PnP Quaternion Solver — 3D spin vector solver outputting frame-by-frame RPM & rotation axis (q). Acceptance: Validated accuracy within ± 3% margin against synthetic & physical drill benchmark datasets.
• Technical Documentation — Clean Jupyter notebooks showing algorithm step-by-step with comments. Acceptance: Fully reproducible execution on sample .mp4 test files.

How to apply

Email your resume to admin@pixactly.ai, with the role title in your subject line.